Nanocoder v1.28.0 - ACP for Zed, slimmer tool surface, session resume, copy command

[will-lamerton]

Nanocoder v1.28.0 is out. Headline additions: an Agent Client Protocol server so editors like Zed can drive Nanocoder natively, a consolidation of the built-in tool surface from 33 tools down to 19 with automatic profile selection, and session resume that replays the full message and tool-call history into the chat view. Underneath: structured tool-result validation, API-reported context usage, a skill linter, two new first-class providers, and a long list of fixes for local-model workflows.

Built by the Nano Collective - a community collective building AI tooling not for profit, but for the community.

Full changelog lives in the repo at https://github.com/Nano-Collective/nanocoder.

Agent Client Protocol (ACP) for Zed and other editors

Nanocoder can now run as an Agent Client Protocol agent:

nanocoder --acp

When launched with --acp, Nanocoder speaks JSON-RPC over stdin/stdout and the editor becomes the UI. Streaming text, tool cards, before/after diffs for string_replace and write_file, permission prompts, model switching, and ask_user options all render in the editor instead of the terminal. The four development modes (normal, auto-accept, yolo, plan) are exposed as session modes selectable from the editor; sessions start in auto-accept.

The new module lives at source/acp/ and covers the agent, server, session, conversation loop, content conversion, capability negotiation, and permission handling. Thanks to @Avtrkrb. Closes #529.

Follow-up work added the missing ACP docs, kept the parallel tool-execution loop in sync with the plain shell, and made NANOCODER_MAX_TURNS configurable (with a raised default) so long ACP and plain-mode runs are not cut short mid-task.

Setup in Zed is two lines in settings.json:

{
  "agent_servers": {
    "Nanocoder": {
      "command": "nanocoder",
      "args": ["--acp"]
    }
  }
}

Full setup notes and limitations (in-memory session history across editor restarts, selection-only ask_user, no image or audio processing) are in the ACP docs at https://github.com/Nano-Collective/nanocoder.

Tool surface consolidated: 33 down to 19

The built-in tool surface has been consolidated from 33 tools down to 19, with a new auto-detected tune profile as the default. The headline shape changes:

• Tasks collapse from four tools into a single write_tasks that replaces the whole list. No IDs for the model to juggle, just a current state.
• File ops merge delete, move, copy, and create_directory into one file_op.
• Git shrinks from eleven tools to six: status, diff, log, add, commit, pr. Rarer operations flow through execute_bash.

On top of the consolidation, the auto tune profile is now the default. It infers full, minimal, or nano from the active model's parameter count, so a small local model automatically gets the slim tool set and shortened prompt while a cloud model is unchanged. The profile re-resolves live on model switch and is surfaced in the input indicator (e.g. tune: nano (auto)). /usage and the context indicator rebuild from the live tune, mode, and model with the profile-filtered tool count, so the numbers reflect what is actually loaded.

Roughly 6.3k lines were removed across this work.

Structured tool results and single-source validation

Tool arguments are now type-checked against each tool's JSON schema at the execution boundary (schema-validate.ts). A malformed model output returns a clear, field-level error the model can self-correct from instead of being coerced at the render layer. Validation runs through a single withValidation seam shared by both the native and XML-fallback execution paths, ahead of any per-tool validator.

Approval policy is centralised the same way, and the old global mode context is removed. The current development mode (including a mid-run switch via Shift+Tab) is honoured consistently across the main loop and subagents. Subagents no longer capture normal once at startup and silently ignore a later switch.

Session resume with full history replay

Resuming a saved session now replays the message and tool-call history into the chat view via a new session-history-renderer, so you pick up exactly where you left off with the conversation visible rather than an empty screen.

Companion fixes resolved an autosave race, a history-truncation edge case, and resume-after-save paths, all with regression coverage. Thanks to @akramcodez. Closes #545.

/copy for the last assistant response

A new /copy slash command copies the most recent assistant response to the system clipboard. Cross-platform via clipboardy (pbcopy on macOS, xclip on Linux, clip.exe on Windows). Surfaces a clear warning when there is no assistant message and a real error when the clipboard write fails.

Safer local-model defaults

Two additions aimed squarely at smaller local models running unattended:

• Read-before-edit guards. string_replace and write_file now require the file to have been read first. A blind edit to a never-read file is rejected at the tool boundary.
• Tool-call loop detection. A tool-signature tracker detects a model repeating the same tool call in a tight loop and breaks the cycle.

Both make local models materially safer to leave running on their own.

Skill linter and command template ergonomics

A new skill linter - /skills check <name> plus a check_skill tool - validates a bundle from disk using the same parsers the loader uses. A PASS means it will load. It is wired into /skills create so the model verifies and self-corrects what it generates. The linter inspects member template bodies too (unsupported mustache tags, unbalanced or unclosed sections, placeholders referencing undeclared parameters), not just frontmatter.

Custom-command bodies now support optional arguments with inline defaults (name=default) and conditional sections using mustache's section syntax (# name for "if name is set" and ^ name for "if name is empty"). Inverted-section support in custom-tool templates goes through a shared section engine, so the same rules apply across primitives.

Two new providers

• Atlas Cloud is now a first-class provider and sponsor. Wizard onboarding template and a dedicated provider docs page ship with the release. Atlas Cloud is a full-modal inference platform that exposes LLM, image, and video generation APIs through a single endpoint.
• Requesty (https://requesty.ai) is added as a first-class OpenAI-compatible provider, mirroring the OpenRouter integration with a name matcher, app-attribution headers, a wizard onboarding template, and docs. Thanks to @Thibaultjaigu. Closes #589.

API-reported context usage

The ctx: NN% indicator now reflects provider-accurate token counts when the model reports them, captured from the final step's usage rather than the tool-loop-summed totalUsage. When the provider does not report, Nanocoder falls back to a client-side estimate, marked with a leading ~ (e.g. ctx: ~42%); API-reported figures render bare.

Follow-up work anchored the API figure across turns, counted real tool definitions in the auto-compact gate, and added a tiktoken-based generic fallback tokenizer for more accurate counts on models without a native tokenizer. Thanks to @JimStenstrom. Refs #381.

Smaller things worth knowing

• Git branch display in the boot summary and /status panel, with narrower-terminal-friendly rendering. Thanks to @ragini-pandey. Closes #539.
• Provider model discovery in the wizard now surfaces real discovery errors instead of silently falling back, and shows model names with long labels truncated by ellipsis. Thanks to @akramcodez and @Dhirenderchoudhary. Closes #554.
• Slash-command completion supports arrow-key navigation and highlighting, with the double-submit-on-select bug fixed (TextInput ignores Enter; UserInput handles select-then-submit behind a guard flag) and customCompletion restored. Thanks to @rakshith1928. Closes #578.
• Force-compact-and-retry fallback. When the LLM server rejects a request for exceeding its context window, Nanocoder now compacts and nudges the conversation to continue instead of failing outright. Thanks to @lordoski. Closes #546.
• Bash executor supports timeout, output limits, and abort, with regression tests covering the executor lifecycle. Thanks to @akramcodez. Closes #547.
• ask_user rendering dropped the leading ?, fixed spacing and the hanging indent, added arrow-key navigation, and normalised object-shaped options to readable strings (preferring a label over a value id).
• @-file-mention handling keeps mention placeholders in the chat instead of expanding inline, and large-file inlining is capped to avoid blowing up the prompt.
• Custom display: failed tool results condense to a one-liner in compact mode (e.g. write_file failed.), while the model still receives the full error in history.

Fixes

• Copilot GPT-5 reasoning streams are now handled correctly via an @ai-sdk/openai patch and chat-handler changes, so reasoning content streams properly instead of breaking the response. Thanks to @EntropyParadigm. Closes #577.
• Propagate AbortSignal to streaming bash execution, so cancelling a turn actually stops a long-running streamed command. Thanks to @Dhirenderchoudhary. Closes #550.
• Corrupted agents.config.json is no longer silently overwritten. The wizard now honours the projectDir prop for config-path resolution and the infinite render loop on the existingProviders default param is resolved, all with regression coverage. Thanks to @Dhirenderchoudhary. Closes #551, #552.
• read_file returns an empty marker instead of throwing on empty files, including range and metadata-only reads. The marker constant is extracted (originally EMPTY_FILE_MARKER, later renamed EMPTY_CONTENT_MARKER) with tests for files containing only a newline. Thanks to @ragini-pandey. Closes #530.
• Compaction no longer splits recent tool-call groups, which previously produced empty model output after compaction.
• / key no longer scrolls the chat view to the bottom during input. Thanks to @itsishant. Closes #590.
• Chokidar watcher is closed if daemon startup throws, and the daemon stop-handler is set up before resources start so a failed lockfile write calls stop() and cleans up. Thanks to @OMEE-Y and @rakshith1928. Closes #553, #557.
• A stray console.log in message-queue.tsx was replaced with the structured logger. Thanks to @A-S-Manoj. Closes #588.
• Subagent context window overrides so delegated agents can run with a different context limit than the main session, and provider names resolve case-insensitively with custom CA bundle support for provider TLS. Thanks to @zerone0x.
• A tmp >=0.2.6 override resolves a path-traversal advisory in a transitive dependency. Thanks to @ragini-pandey.

Refactor and dead-code sweep

Beyond the tool-surface consolidation above: unified tool-call ID generation, extracted a shared useWizardForm hook, consolidated session-override managers, added message-factory helpers, deduped config loaders, git exec, command dispatch, and conversation-loop flush, extracted StyledSelectInput and makeSimpleToolFormatter, and deleted several dead modules and orphaned exports (including fetch-local-models, orphaned by an earlier removal).

The Nanocoder Battlemap competitive comparison and the README and docs have been refreshed.

Dependency updates

• ai 6.0.174 to 6.0.193, plus @ai-sdk/openai, @ai-sdk/openai-compatible, @ai-sdk/anthropic
• @agentclientprotocol/sdk 0.22.1 to 0.25.0
• undici, esbuild
• @biomejs/biome 2.5.0
• lint-staged 17
• @types/node, @types/vscode, and other transitive bumps tracked through the lockfile
• clipboardy ^5.3.1 for /copy

If anything is broken or surprising, please open an issue or drop a note in Discord. Thank you for using Nanocoder.